© Copyright JASSS

  JASSS logo ----

Armano Srbljinovic, Drazen Penzar, Petra Rodik and Kruno Kardov (2003)

An Agent-Based Model of Ethnic Mobilisation

Journal of Artificial Societies and Social Simulation vol. 6, no. 1

To cite articles published in the Journal of Artificial Societies and Social Simulation, please reference the above information and include paragraph numbers if necessary

Received: 16-Nov-2002      Accepted: 15-Jan-2003      Published: 31-Jan-2003

* Abstract

In this paper we used the methodology of agent-based modelling to help explaining why populations with very similar socio-demographic characteristics sometimes exhibit great differences in ethnic mobilisation levels during mobilisation processes. This agent-based model of ethnic mobilisation was inspired and developed by combining and extending several theories, ideas and modelling constructs that were already used in agent-based modelling of social processes. The model has been specifically adapted to account for some of the most important characteristics of ethnic mobilisation processes that took place in the former Yugoslavia. Results obtained by experimenting with the model indicate that the observed differences in mobilisation levels across populations may sometimes not be related to the variations within any particular socio-demographic factor, but merely to random differences in the initial states of the individuals. In this model these random differences primarily relate to the degrees of importance that individuals attach to their ethnic identity, as well as to the layout of social networks.

Agent-Based Modelling; Ethnic Identity; Ethnic Mobilisation

* Introduction

Recent conflicts in the former Yugoslavia attained a multitude of forms: from political confrontations to a full-fledged war; from divides between regions, villages, city blocks, neighbours and family members to inter-national and inter-state conflicts; from events characterised by sequences of rationally calculated moves to the examples of individual and massive psychoses; from the actions motivated by the sense of loyalty or responsibility to a particular group or humanity in general to the moves motivated by greed or will for domination; from the examples of conflict resolution reached by an agreement to ethnic cleansing and genocide. These provoke the question of which processes contribute to the escalation of social conflict and organised violence and, if possible, how to influence those mechanisms in order to avoid instability in a society, region, or perhaps the whole world.

Recognising that the key to answering such questions is better understanding of social facts and processes related to the roots, the nature, the course of events and the consequences of recent conflicts, the Sociology Department of the Faculty of Philosophy at Zagreb University initiated the scientific project called "Social Correlates of the Homeland War". One of its aims is the development of formal models that could be used as tools in discovering, investigating and explaining the social correlates of recent conflicts. In doing so, the intention is also to exploit the capabilities of modern computational technology.

This paper describes a computational model of ethnic mobilisation that is being developed as part of this project, in the co-operation between the Department and the Croatian MoD's Institute for Defence Studies, Research and Development. The aim is to build the model of ethnic mobilisation based, as much as possible on the contemporary scientific theory, and capturing, with at least qualitative fidelity, some of the most important characteristics of ethnic mobilisation processes that took place in the former Yugoslavia. The methodology of agent-based modelling is used to help explaining why populations with very similar socio-demographic characteristics sometimes exhibit great differences in ethnic mobilisation levels during mobilisation processes.

The main purpose of this modelling effort is, to borrow from Forbes, " […] to express as straightforwardly as possible a few key ideas, so as to bring out more clearly the logic of the reasoning, in particular the assumptions being made about the interdependence of […] distinct causal processes […]." (Forbes 1997, pp. 165-166) By exploring the consequences of such assumptions in an artificially constructed, highly simplified society, we do not have an (possibly unattainable) aim of replicating the real-world processes exactly or determining the exact values of the model's parameters. Our more moderate goals are based on a seemingly reasonable assumption that the results observed in a simplified, artificial society could give us some clues of what is going on, or perhaps show us where to centre our attention in further and more detailed examination of a more complex real-world society.

The paper is organised as follows. Section 2 describes the context of the modelled situation. Section 3 describes the model, first in general and then in more details. Section 4 presents the experiments done with the model so far and discusses the main results that were obtained. Section 5 outlines several possible routes for further exploration.

* The Model's Wider Context

The term ethnic mobilisation, as used in this paper, does not denote the transition from an ethnic group unconscious of its own cultural specifics to the one which is conscious of them, neither does this term here denote the transformation of an ethnic community lacking political organisation into an ethnic community possessing such an organisation. This paper concentrates on the process of ethnic mobilisation within a multinational state, comprised of already formed nations. These nations are conscious of their own cultural specifics and have considerable political rights within the multinational state system, but they have not got their own fully sovereign states. The term ethnic mobilisation, as used here, refers to the process of reviving their latent ethnic identities. More specifically, we investigate what has already been described as social situations in which ethnic roles, that used to have less social importance, are, under certain circumstances, pushed up towards the highest importance end of the scale (Banton 1994).

The above-mentioned situation existed in the former Yugoslavia, which, according to its 1974 Constitution, consisted of six republics. The majority of the population in each republic comprised one of Yugoslavia's constituent nations[1]: Muslims in Bosnia and Herzegovina, Croats in Croatia, Macedonians in Macedonia, Montenegrins in Montenegro, Serbs in Serbia, and Slovenes in Slovenia. The territories of the republics corresponded to the historical, political, geographical and cultural wholes established earlier, but this also implied that the republic borders were not always concurrent with the ethnic ones.

By the end of the eighties and the beginning of the nineties the whole region experienced a strong ethnic revival which resulted in a violent collapse of the federal state and the formation of new internationally recognised states that correspond to the former Yugoslav republics.[2] Avoiding detailed examination of causes of such a collapse, we just mention that one can find several explanations in literature, differing in the amount of importance that they ascribe to different factors.[3] Among the most prominent and widely cited are instrumental explanations emphasising the role of political entrepreneurs - nationalist elites and leaders like Slobodan Milosevic, who consciously mobilised masses and acted on Yugoslavia's dissolution in order to gain political advantages of their own (Silber and Little 1997). Although these explanations are mostly correct, they are not complete in the sense that they raise the question of why were those entrepreneurs so 'successful', and why were they 'successful' at that particular time. The proponents of explanations based on the importance of particular political processes and institutions that existed in the former Yugoslavia argue in favour of such factors as those that enabled 'success' of political entrepreneurs, pointing out also the role of the wider economic and international situation at the end of the eighties and the beginning of the nineties (Woodward 1995; Crawford 1998). These explanations, however, raise further question of why particular political, economic and other institutions of the former Yugoslavia, that were in favourable international conditions so easily misused or transformed by nationalist leaders, came into existence in the first place. Namely, it can be argued that the former Yugoslavia's founders did not have many choices when designing the federal state institutions that could keep together several already existing nations with rather long histories, conscious of their cultural specifics, already possessing considerable political rights within the earlier state systems existing in the region, and burdened already with their collective memories of mutual disputes and mistrust. Taking this into account, we can say that the explanations emphasising deeper societal conditions such as history and culture, as well as the emotional power of ethnic identity sentiments are also correct, at least in a degree in which we do not articulate them as mere 'stories of ancient Balkan hatreds', but as attempts of explaining existing strong inter-national tensions in the light of the region's history.

Such discussion concerning the causes of Yugoslavia's collapse can also be summarised in terms of the framework introduced by Cederman (2001a). He distinguishes three approaches to the relationship between culture and politics: the essentialist approach - emphasising the role of pre-existing cultural 'raw material' on the articulation of political identities, the instrumentalist-constructivist approach - emphasising the role of political entrepreneurs in the formation of a political identity with culture as a mere side-effect, and the limited-constructivist approach that complements the instrumental logic with an institutional feedback. According to this last view, political entrepreneurs do shape 'cultural material', but their freedom of choice is limited, because, once present, cultural and ethnic boundaries get institutionalised and acquire an autonomous role feeding back into the political process. Reflecting on the previous discussion, we may say that this third approach was the one that led us in constructing our model.

So, by the end of the eighties and the beginning of the nineties, the ethnic roles in the society of the former Yugoslavia, that were kept toward the middle of Banton's social roles-scale for more than forty years, now under the influence of political entrepreneurs, increased in importance. As Banton observes: "It appears that […] some people have been forced […] to regard their ethnic identity as more important relative to other identities than they used to do. Whereas previously they could have placed obligations to a friend, neighbour or fellow worker before an obligation of shared ethnicity alone, they have been forced by processes which we do not fully understand, to change their priorities." (Banton 1994)

The aim of the model that will be described next is to contribute to a better understanding of the processes to which Banton refers.[4] In doing so, our further intention is to account for the fact that often geographical areas with very similar socio-demographic characteristics exhibited great differences with respect to their populations' mobilisation levels during mobilisation processes. For example, while certain rural areas in the Croatian region of Lika with mixed Croatian and Serbian population became foci of ethnic conflicts, other, geographically close, rural and similarly ethnically mixed areas of the Gorski kotar region remained stable throughout the entire conflict period, although they were equally accessible to political entrepreneurs.

* The Model

The agent-based model that we propose here is inspired by already existing theories of mobilisation (Szayna 2000; Moore and Jaggers 1990). It has been developed by combining and extending several ideas and modelling constructs that had already been used in the agent-based modelling of social processes (Axelrod 1997; Cederman 1997; Lustick 2000; Hammond 2000). The model has also been specifically adapted to account for some of the particularities of the situation described in the previous section. Our hope was that by using an agent-based modelling method we could obtain some clues as to how it is possible that a variance appears in mobilisation levels across similar populations. In this section we will first broadly explain the main model's characteristics and relationships, and then more closely describe how these general ideas are actually implemented.

As we have already noted, the situation to be modelled is the one in which nations with formed national identities already exist, but in which the importance they attach to those identities varies with time. To account for this fact we provided our agents, who are meant to represent individual members of a society, with two basic attributes. First, they have 'ethnic membership', the 'red' or the 'blue' one[5], which is fixed in all simulation runs. Second, agents possess a degree of ethnic mobilisation - the degree to which they identify themselves with their ethnic group or the degree to which they attach importance to their ethnic identity. This can vary between 0 and 1, so that in this model the identity we are talking about is, so to say, 'a matter of degree'. To express the earlier mentioned tension between ethnic and other possible identities, we interpret the difference between the maximum of 1 and the actual degree of agent's ethnic mobilisation as the degree of importance that an agent attaches to his other ('non-ethnic') acceptable identities - professional, regional, etc. This can also be interpreted as a degree of 'civic mobilisation' - the degree to which an agent is sensitised to the inclusiveness in a civic society. Thus, increasing the degree of ethnic mobilisation implies decreasing of the degree of civic mobilisation and vice versa, as seems to be the case in reality.[6]

Modelled agents also possess an attribute called 'grievance degree', supposed to capture agents' (dis)satisfaction with their life conditions: economic, political, security and other. The grievance degree can also attain values between 0 and 1. In accordance with a widely accepted view (e.g. Gurr 1998), grievance intensities are supposed to facilitate the mobilisation process.[7]

In addition, each agent possesses its own 'social network' representing other agents with whom it communicates (family, friends, etc.). Therefore the agent can observe their identities and degrees of mobilisation and grievance. On the basis of such observations agents perceive their 'environmental conditions', which influence the state of their variables. Members of social networks are for each agent, by default, chosen randomly among the members of the population.

During the simulation agents receive 'appeals'. The mechanism of appeals is one of the primary means in altering preference rankings of a targeted population during mobilisation processes (Moore and Jaggers 1990). In our model appeals are meant to represent various means - media, public meetings, etc. - by which particular political entrepreneurs - ethnic leaders, political organisations, state governments, non-governmental organisations, etc. - influence individuals' ethnic/civic orientation. All these means of issuing appeals were present in the case of the former Yugoslavia.[8]

Appeals are characterised by their source, which may be blue, red or grey, i.e. neutral; and by their content, which may be 'increase' or 'decrease mobilisation level'. 'Coloured' appeals increase, while neutral appeals decrease agents' mobilisation levels. In accordance with the adopted limited-constructivist approach, the entrepreneurs' appeals are allowed to influence only "the volume of agents' radio-stations", but not to "change the stations at which agents play". This is not to say that we do not accept that ethnic and national identities can change under the influence of entrepreneurs, but that this is hardly feasible in a characteristic time-span - several years or so - that the model refers to, particularly so in the specific, previously described social situation.

In the absence of appeals, agents' mobilisation levels gradually decrease. This is justified by the fact that high mobilisation levels require investment of energy and resources, so they tend to decrease without external influences.

Formalising the above outlined general ideas, we came up with the following expression describing the dynamics of mobilisation intensity of agent i:

mi(t+1) = mi(t) + (miapp + misocnet + micoolt,(1)
where mi(t) is the mobilisation intensity of agent i in time t, miapp is the change of mobilisation intensity of agent i due to appeals received between times t and t+1, misocnet is the change of mobilisation intensity of agent i due to the influence of its social network, and micool ≤ 0 is the change of mobilisation intensity of agent i due to 'cooling effect'.

The expression describing the change of mobilisation intensity of agent i due to received appeals is taken to be as follows:

Eqn 2

in which gri denotes the grievance degree of agent i,[9] and kapp is a constant, which together with the values of the second coefficient ksame/other/neutral, enables us to control the magnitude of miapp, i.e. to control agent i's 'susceptibility to appeals'. As we can see in (2), we distinguish between agents' susceptibilities to appeals issued from the sources of the 'same', the 'other', and the 'grey' colour. It is generally assumed that, all other things being equal, the effect of an appeal is stronger on the agents of the same colour as the appeal's source, than on the agents of other colour. Consequently, the values of ksame are greater than the values of kother. It is also assumed that the effect of 'a neutral appeal' is generally smaller than the effect of 'brethren's appeals', but still stronger than the effect of 'appeals of the others'. Finally, regarding the effects of neutral appeals, it is assumed that these effects are proportional to the intensity of civic (not the ethnic) mobilisation, so the factor mi(t) is in the case of neutral appeals replaced with 1 - mi(t).

The expression describing the change of mobilisation intensity of agent i due to the influence of agent's social network is the following one:[10]

Eqn 3

in which ksocne is the coefficient controlling the magnitude of misocnet, netsize is the size of the network, i.e. the number of agents comprising the social network of agent i, and impsame, impother are impacts coming from the agents of the same and other colour, respectively. More precisely, the formula for the impact of those agents in the network that have the same colour as agent i is:

Eqn 4

in which Nsame is the number of such agents. The formula for the impact of those agents in the network that have colour different than the colour of agent i is:

Eqn 5

in which Nother is the number of such agents.

The first term in both formulas is the actual difference in mobilisation intensities between agent j in the social network of agent i and the agent i itself. This term is multiplied with another term representing 'the minimal common ground for communication' that the two agents possess. Namely, we assume that the social influence between two agents of the same colour exists in both ethnic and civic dimensions. Therefore 'the minimal common ground' for two agents of the same colour is a combination[11] of their 'minimal common ground' in the ethnic dimension; and their 'minimal common ground' in the civic dimension. Their 'minimal common ground' in the ethnic dimension is represented by the lesser of their ethnic mobilisation identities, while their 'minimal common ground' in the civic dimension is represented by the lesser of their civic mobilisation identities. Similarly, we assume that the social influence between two agents of different colours exists in civic dimension only, so 'the minimal common ground' for such agents consists of the lesser of their civic mobilisation identities. The main desired effect of this construction is that for two agents j1 and j2, whose current differences in mobilisation levels with respect to agent i are equal, the higher influence on agent i will be the agent whose 'minimal common ground' with respect to i is 'broader'.

Finally, the expression describing 'the cooling effect' is as follows:

Eqn 6

in which kcool is the coefficient controlling the magnitude of micool. This ensures that cooling increases exponentially with mobilisation intensity, that it is very small for small mobilisation intensities, and that it is zero for the mobilisation intensity of zero.[12]

* Experiments and results

The model is implemented using the Java version of SWARM programming package (version 2.1.1) (http://www.swarm.org/) on a HP Kayak XU PC Workstation with the MS Windows NT operating system. All the results presented here have been obtained using 200 agents: one hundred red ones and one hundred blue ones.[13] If not stated otherwise, the following default parameter values are assumed: k1 = 0.1, ksame = 1.0, kother = 0.25, kneutral = 0.5, ksocnet = 0.25, kcool =0.01, netsize = 6. In addition, the default setting assumes uniform distribution of agents' initial mobilisation intensity on the interval [0.0, 1.0], constant values of agents' grievance degrees of 0.5,[14] and random assignment of the members of agents' social networks [15][16].

The appeals are sent to the entire population in regular time intervals, the length of which[17] is specified by the user. For simplicity, at this research stage we assume that all appeals reach all population members.[18]

To facilitate the inspection of model's behaviour during the very first simulation experiments, we have used a typical SWARM-based graphical user interface, where agents are visualised as square cells, coloured in red or blue, so that the colour depicts agents' ethnic membership. The intensity of redness or blueness represents the intensity of agents' ethnic mobilisation. To facilitate comparisons, we have also grouped the agents appropriately, so that the red ones occupied one, and the blue ones the other half of the board (Figure 1).[19]

Fig 1
Figure 1. Visualisation of agents using a SWARM-based GUI

Starting with the experiments we have first verified that the 'simple behaviours' that were anticipated when building the model were as expected. For example, and the most trivially, increasing the frequency of coloured appeals increases the speed of the mobilisation process, while increasing the frequency of neutral appeals accelerates demobilisation. Furthermore, perhaps a little less obviously, mobilisation is faster when agents possess social networks, than when they do not.[20] When building the model, we could only hope that the effect of the average number of 'friends' of the other colour on slowing the mobilisation process would be observable, reflecting the greater 'ethnic tolerance' of such agents - and this effect really appeared.

During these initial explorations we have noticed that for most combinations of appeal frequencies, the outcomes of simulation runs are relatively straightforward. Namely, as one would also expect, populations - the red and the blue one - reach, after a certain simulation time, the extremal mobilisation levels of zero or one, depending on whether the mobilisation or the 'cooling' tendency[21] for the population is stronger. However, for some particular values of appeal frequencies, the average mobilisation levels of the two populations remained approximately constant[22] for a hundred or even more simulation periods. We decided to take a closer look into these cases where the two opposite tendencies seemed to be 'in balance', feeling that those cases could be illustrative for the phenomenon we wanted to explore.

As this was observed in the default setting with the frequency of red appeals of 3 [simulation periods] and the frequency of neutral appeals of 4, we decided to repeat simulation runs in this setting hundred times, each with its own, randomly chosen random generator's seed, and each proceeding for a thousand simulation periods. The outcomes were quite surprising. More than 60% of the runs eventually resulted with the well-known outcomes of both populations reaching zero-mobilisation, or with the 'reds', as primary appeal targets, reaching one, and the 'blues' reaching zero. However, in some cases this final state occurred only after initial oscillations. Even more interestingly, there also appeared outcomes with stable states other than zero or one, as well as the oscillatory outcomes without any stable state at all (Figure 2). Let us remind once again that this variety of outcomes is produced by variation in initial conditions only, as all other model's parameters are kept constant. Moreover, the global statistics of the initial setting are also constant, with only agents' local initial conditions - mobilisation intensity and composition of social networks - being variable.

Fig 2
Figure 2. Some of the typical outcomes produced by variation in agents' initial mobilisation intensity and in agents' social connections, at the frequency of red appeals 3 and the frequency of neutral appeals 4

Next, we decided to investigate whether the composition of agents' social networks will have any impact on the observed outcomes, so we repeated this experiment, but with varying the values of probability p of having a friend of a different colour. Table 1 summarises the results. We classified types of outcomes into 9 broad categories according to similarity, and counted the number of occurrences of each outcome in hundred runs with various probabilities. Interestingly, it appeared that for relatively homogeneous networks,[23] only the two 'simple' outcomes appeared: 'both reds and blues to zero' (type I) and 'reds to one, blues to zero' (type II). However, with the increase in networks' heterogeneity, the diversity of outcomes increases as well (types III to IX - various oscillatory forms), being maximal at the value 0.5 for the probability of having a friend of a different colour. Increasing that value further, diversity of outcomes diminishes again.[24]

Table 1: Distribution of outcomes according to their types, when varying the value of probability p of having a friend of a different colour

pNumber of outcomes of type:

The next series of simulation runs was very similar to the previous one, except that we dispersed grievance degrees, which were previously kept constant at the value of 0.5.[25] We used uniform distribution of grievance with the mean value of 0.5, and varied its variance. Several new kinds of outcomes appeared - the 'heartbeat'-pattern being among the most peculiar ones (Figure 3, upper right part). Generally, however, with increasing the variance, diversity of outcomes diminished. The most frequent outcome appeared to be stabilising the average red mobilisation intensity around 0.9, and the average blue mobilisation intensity below 0.5 (Figure 3, upper left part). This outcome resulted in approximately 80% of simulation runs when grievance was uniformly distributed on the interval [0.0, 1.0].

Fig 3
Figure 3. Some of the patterns observed when dispersing grievance values

We have also started experimenting with cases when both red and blue appeals are sent to the population. Although such cases also proved to be sensitive to variation in the initial conditions, we have not detected particular frequency combinations for which small random variations in initial conditions would produce a great variety of different outcomes as in the above cases with red and neutral appeals only. However, it seems to be too early to draw conclusions regarding this observation, as more thorough explorations are yet to come. As for now, we note that in one hundred simulation runs in our default setting, with the frequencies of both red and blue appeals of 6 and the frequency of neutral appeals of 4, in approximately 50% of cases the outcome is 'both reds and blues to one', the remaining 50% being 'both to zero'.

Finally, we have also explored what happens when introducing blue appeals in the setting described earlier with the frequency of red appeals 3 and the frequency of neutral appeals 4. We repeated one hundred runs with exactly the same seeds as before, but in the presence of blue appeals with various issuing frequencies. An interesting observation is that the introduction of blue appeals, with parameters as in the initial setting, actually 'helps the red side'. For example, when the frequency of blue appeals is 200, 15 out of hundred outcomes are changed, most of them ending with average red mobilisation of one, and the blue one of zero (Figure 4). At the blue frequency 50, the same outcome results in 89 out of 100 cases. However, preliminary experiments suggest that decreasing the value of agents' susceptibility to appeals of the other colour kother may reverse this trend in favour of the blue side.

Fig 4
Figure 4. The example of change of an outcome with the change in blue frequency from none (left) to 200 (right) (default value kother = 0.25)

Some of the presented results are rather tempting in the sense that they are quite readily interpretable. For example, it is tempting to assert that the experiments with variable composition of social networks indicate that greater contact leads to greater uncertainty in outcomes. On the other hand, some results do not seem easily comprehensible and seek explanation. Taking this into account, together with the fact that we have not yet explored the model's parameter space to the full, we would like to refrain from more detailed interpretations of the model's behaviour at this research stage.

However, the fact is that the 'simple model's behaviours', as we called them, are perfectly intuitive and that the model captures all the simple phenomena that it was intended to capture. Moreover, it does so using relatively simple and, if not fully theoretically grounded, then at least plausible modelling constructs. This gives us enough confidence to say that, if nothing else, more complex behaviours produced with such mechanisms deserve our attention. Namely, if simple modelling constructs and mechanisms may produce complex and sometimes hardly anticipatable behaviours in an artificial setting, then the similar, or even more complex mechanisms at work in the real world, may produce only comparably, if not even more complex behaviours, but most probably not the simpler ones.

Of particular significance is that we also succeeded in generating differential mobilisation levels across similar populations. More precisely, the model indicates that differences in mobilisation levels across populations may sometimes not be related to variation in any particular socio-demographic factor, but merely to random differences in the initial degrees of importance that individuals attach to their ethnic identity, as well as to random differences in the layout of social networks.

The observed model's high sensitivity on random variations in the initial distribution of mobilisation intensity and social networks may indicate inherent limitations on the predictability of mobilisation processes. The question that almost inevitably ensues is what is the use of these observations. Namely, if agent-based models or any other tools help us discover that some social processes such as ethnic mobilisation, once started, are hardly predictable and controllable, how could this finding possibly help us? Would it merely mean that we have to abandon further investigation of such processes? The answer is negative, because even realizing that a process is barely predictable or controllable is better than no understanding at all, and may still help preventing the undesirable effects of the process. Such an understanding may stimulate us, for example, to become more sensitive to certain processes, or more inclined to avoid them knowing their potential dangers. Viewed in this light, preliminary results obtained with the model need not be taken as pessimistic or depressing. On the contrary, they may serve as a basis for further investigations.

* Further work and possible extensions

What is needed next is to explore the capabilities of this model to the full. What happens for various combinations of appeal frequencies when varying initial mobilisation and grievance conditions? What if appeals are sent selectively instead of to the whole population? What if they are sent periodically or randomly in time? How do the results depend on population size? These questions still wait to be answered.

Then comes the question of modifying particular mechanisms used in the model. For example, grievance levels could change, like mobilisation levels already do, under influences coming from social networks, or under 'external influences' - representing economic, political, military and other sets of measures that may be employed in moderating or raising grievance levels.

In fact, the relationship between mobilisation and grievance is probably much more subtle than reflected in the model, as high levels of mobilisation often produce collective actions, which may endanger society's economy and security and spread discontent among the population, particularly in the long run. Increased mobilisation in the short run, on the other hand, may well serve as a kind of emotional outlet, and actually decrease the grievance levels. However, the implementation of such relationships might easily complicate the model beyond reasonable limits. As we have also developed another model representing society on a more aggregated level, and focused more on the collective action issues (Penzar and Srbljinovic 2002), we are considering the possibility of integrating these two models in future.

Another rather unrealistic assumption, introduced mainly to keep the model as simple as possible at this research stage, is that agents' social networks are static, in the sense that social networks' layout does not change with changes in agents' mobilisation levels. A possible extension of the model could attempt to rectify this issue. Insights obtained from belief formation and information diffusion theories, as well as studies of the former Yugoslavia's social networks system, could be used to further improve the model's network representations.

It is also interesting to observe that the model presented has some characteristics in common with the so-called sand-pile models. In such models steady, linear input - similar to the appeals-mechanism of this model - generates tensions inside a system that lead to non-linear and delayed output ranging from small events or 'avalanches' to huge ones. The sizes of the output events are distributed according to a 'power law' with the size of an event inversely proportional to its frequency (Bak 1996). Recent research results show that the distribution of nationalist wars also conforms to the power law (Cederman 2001b), which may well indicate that the process of ethnic mobilisation should be modelled using models of the sand-pile type. This hypothesis is further justified by frequent accounts of ethnic mobilisation campaigns where such campaigns are likened to "the spread of fire", and, as is already known, forest fires also follow the power law distribution (Turcotte 1999). Thus, another promising route for future research could be investigating possible extensions of the model presented here that could generate power-law distributed 'avalanches' of ethnic mobilisation.

Besides rather abstract model settings discussed up to this point, we hope that model settings that would roughly correspond to select mobilisation cases from the former Yugoslavia, could be set up and run for closer examination, enabling the model's validation, at least to some extent. The problem aggravating this issue is the lack of reliable data concerning the dynamics of mobilisation of ethnic groups in former Yugoslavia, let alone the more general problem of measuring ethnic identities at all (Posner 2000).

With only illustrative purpose, we may just briefly reflect on some of the calibration and measurement scale problems that are to be expected. This model operationalises concepts like mobilisation and grievance as ratio scale variables. On the other hand, it is hard to expect that any kind of archival or similar studies, which could be used for validation purposes, would yield mobilisation data on a scale higher than ordinal. The immediate consequence, of course, is that only partial validation would be possible. It is also to be expected that the time resolution of observed changes in the mobilisation variable levels would be closer to months or quarters of a year, than to hours or days. Taking into account that our model produces 'a more continuous', ratio scale output, its parameters would probably need to be calibrated so that the simulation periods are closer to days or weeks, in order to allow such an output to be compared to the ordinal and time-discrete data.[26]

We also hope that the model presented here is a small step toward the creation of a larger number of models that would eventually enable better understanding of the recent events in the former Yugoslavia, and better use of the accumulated historical experience. While not disregarding how unpleasant and traumatic this experience may be, we feel that the best we can do with it is to use it for the purpose of improving our knowledge.

* Acknowledgements

Originally prepared for presentation at UCLA Computational Social Sciences Conference, Lake Arrowhead, California, May 9-12, 2002. The authors would like to express their gratitude to Professors Ozren Zunec, Ph.D. and Vjekoslav Afric, Ph.D., Sociology Department of the Faculty of Philosophy, University of Zagreb, as well as to Brigadier Vjekoslav Stojkovic, Ph. D., Deputy Director of the Institute for Defence Studies, Research and Development, Ministry of Defence of the Republic of Croatia, both of whom provided continual support for this research. A special thanks also goes to Mr. Ognjen Skunca, M.Sc. for introducing the authors to the world of agent-based models.

* Notes

1 Strictly speaking these 'nations' were 'nationalities', because they shared the same internationally recognised state - the former Yugoslavia, but the Yugoslav Constitution called them 'nations', while the term 'nationality' was reserved for national minorities in former Yugoslavia that were not constituent nations: Albanians, Czechs, Hungarians, Italians, Slovaks etc. The Constitution guaranteed all the republics wide autonomy and political rights, including the right to self-determination, which later presented the legal basis for the opinions of the Badinter Arbitration Committee that led to international recognition of those republics as independent states (Pellet 1992).

To avoid this confusion with 'nations' and 'nationalities', and taking into account that the (non)possession of political rights is not of particular importance for our model, in the remainder of this paper we will mostly use the more basic term 'ethnic'. By using this term, we do not mean that nations of the former Yugoslavia are mere ethnic groups and not nations, neither that they are nations less than other nations are. On the contrary, by using this term we want to emphasise that ethnicity is a universal human characteristic.

2 Except for the fact that Montenegro and Serbia stayed together in the same state, the status of which is currently in the process of redefinition.

3 In discussing four distinct classes of 'causes of conflict': actions of political entrepreneurs, political processes and institutions, international influences, and broader societal conditions, we closely follow Lund (2001). We also agree that: "The point of laying out these four types of explanation is not that one or another approach will necessarily be the single correct one. It is rather that in approaching a given conflict, analysts […] may need to consider several kinds of variables." (ibid., p. 135). It is also likely that not all the conflicts in the region of former Yugoslavia were equally influenced by particular factors. An enquiry with the scope much broader than ours would be needed to determine the importance of each group of factors in each particular case.

4 To use Posner's metaphor (Posner 2000), the problem that Banton refers to may be likened to tuning stations on a radio device and may be further broken into two components: first, changing the radio station, i.e. choosing particular (ethnic) identity among possible others, and second, turning up the volume at which particular station plays, i.e. increasing the degree of importance of that identity. This work deals mainly with the second component of this problem.

5 Some cases of ethnic mobilisation in the former Yugoslavia, particularly in Bosnia and Herzegovina, involved simultaneous mobilisation of more than two different ethnic groups. However, as we shall see, even the two-sides model becomes soon very complex, so, for the sake of simplicity, we have not considered modelling of more than two groups so far.

6 Having only two 'rival attributes' is also a simplification. In the particular case of former Yugoslavia, other attributes existed, like religion and membership in the Yugoslav communist party. Religious affiliation (to Islamic, Orthodox and Catholic church) was consistent with ethnic divides and generally played a mobilising role. Membership in a unified Yugoslav communist party spread over all nations and republics, but did not play a moderating role to the conflict. The party instantly broke down along ethnic borders like all other federal institutions.

7 The discussion of a relationship between grievance and mobilisation intensities is continuing in Section5.

8 For a particularly vivid account of the so-called 'Meetings of Truth' accompanying Slobodan Milosevic's ethnic mobilisation campaign that he called 'Anti-Bureaucratic Revolution' see (Silber and Little 1997, pp. 58-69).

9 For the sake of model's simplicity, at this research stage we hold grievance degrees constant during simulation experiments.

10 For the sake of model's simplicity, at this research stage agents use only information about colour and mobilisation intensity of the members of their social network; information about grievance degrees is not used.

11 More precisely, we use the sum of the two. We have done some experiments with the maximum of the two instead of the sum, but have not observed significant differences in results.

12 The choice of this particular 'cooling function' is rather arbitrary. We were primarily led by a desire to keep it relatively simple. Possible model's extensions, discussed in the last section of this paper, may also include experimenting with other functional forms.

13 Choosing the relatively small total number of agents was mainly dictated by the limited computing power at our disposal (Pentium II on 266 MHz).

14 In order to start with as simple model as possible, we chose not to disperse the grievance variable as default.

15 In fact, as we can control the probability with which agents possess friends of a different colour, the 'random assignment' here means that the members of social networks are randomly chosen, and that this is done so that the probability of having a 'friend' of other colour is 0.5.

16 Note that there are two sources of randomness in the default setting: the initial mobilisation intensity and the layout of social networks.

17 I.e. the frequency (in simulation periods) with which the appeals are issued.

18 This 'perfect observability of appeals' can be justified on the grounds that in modern communications era it becomes increasingly difficult to 'hide' mobilising appeals to the own group from 'others'.

19 It is important to observe that this grouping does not reflect agents' actual positions in social networks, i.e. 'geographic proximity' in this representation does not imply membership in neighbours' social networks. The only purpose of the grouping was to facilitate visual inspection of changing mobilisation levels in time.

20 On the other hand, varying the size of the networks did not produce significant effects, which is also easily justified by the fact that the network's influence is normalised with network's size in (4).

21 Reflected in frequency values of issuing coloured and neutral appeals, respectively.

22 I.e., close to the default initial average mobilisation of 0.5.

23 With probabilities of having a friend of a different colour 0.1, 0.2 and 0.3.

24 Settings with probabilities of having a friend of a different colour greater than 0.5 are rather artificial in the sense that such an ethnic group could probably not exist in reality in a long term, as contacts among its members would be too rare to enable sustainability of the group's identity. This is also the reason why we did not experiment with the most extreme values of p: 0.8 and 0.9. However, more rigorous sensitivity analyses in the future may include those experimental settings also.

25 Probability of having a friend of a different colour was again on default value of 0.5

26 There seem to be no obstacles for the simulation time periods to be interpreted as needed: as hours, days, weeks, months or whatever. Determining the right time scale depends mainly on how fast are the real mobilisation processes that we want to refer to. The model seems also to be general enough to allow similar scalability in spatial dimension, i.e. in the number of represented agents.

* References

AXELROD R (1997) The Complexity of Cooperation: Agent-Based Models of Competition and Collaboration. Princeton: Princeton University Press.

BAK P (1996) How Nature Works: The Science of Self-Organized Criticality. New York: Copernicus.

BANTON M (1994) Modelling Ethnic and National Relations. Ethnic and Racial Studies, 17(1), pp. 1-18.

CEDERMAN L E (1997) Emergent Actors in World Politics: How States & Nations Develop & Dissolve. Princeton: Princeton University Press.

CEDERMAN L E (2001a) Nationalism and Bounded Integration: What It Would Take to Construct a European Demos. European Journal of International Relations, 7(2), pp. 139-174.

CEDERMAN L E (2001b) Nationalist Systems Change and its Geographic Consequences. Paper prepared for the Annual Meeting of the American Political Science Association, San Francisco,http://www.sas.upenn.edu/brownecip/Cederman.doc.

CRAWFORD B (1998) "Explaining Cultural Conflict in Ex-Yugoslavia: Institutional Weakness, Economic Crisis, and Identity Politics". In Crawford B, Lipschutz R D (Eds.), The Myth of "Ethnic Conflict": Politics, Economics and "Cultural Violence". Berkeley: University of California.

FORBES H D (1997) Ethnic Conflict: Commerce, Culture and the Contact Hypothesis. New Haven and London: Yale University Press.

GURR T R (1998) "A Risk Assessment Model of Ethnopolitical Rebellion". In Davies J L, Gurr T R (Eds.), Preventive Measures: Building Risk Assessment and Crisis Early Warning Systems. Lanham: Rowman & Littlefield Publishers, Inc.

HAMMOND R (2000) Endogenous Transition Dynamics in Corruption: An Agent-Based Computer Model. Center on Social and Economic Dynamics working paper,http://www.brook.edu/es/dynamics/papers/ross/ross.pdf.

LUND M S (2001) "Why Are Some Ethnic Disputes Settled Peacefully, While Others Become Violent? Comparing Slovakia, Macedonia , and Kosovo". In Alker H R, Gurr T R, Rupesinghe K (Eds.), Journeys Through Conflict - Narratives And Lessons, Lanham: Rowman & Littlefield Publishers, Inc.

LUSTICK I S (2000) Agent-Based Modelling of Collective Identity: Testing Constructivist Theory. Journal of Artificial Societies and Social Simulation, vol. 3, no. 1,http://jasss.soc.surrey.ac.uk/3/1/1.html.

MOORE W H and Jaggers K (1990) Deprivation, Mobilization and the State: A Synthetic Model of Rebellion. Journal of Developing Societies, 6(1), pp. 17-36,http://garnet.acns.fsu.edu/~whmoore/researchpapers.html.

PELLET A (1992) The Opinions of the Badinter Arbitration Committee: A Second Breath for the Self-Determination of Peoples. European Journal of International Law, 3(1), pp. 178-185.

PENZAR D and Srbljinovic A (2002) Dynamic Modeling of Ethnic Conflicts. Paper presented at the 16th triennial conference of the International Federation of Operational Research Societies, University of Edinburgh, July 8-12,http://www.ffzg.hr/hsd/polemos/modeling-ethnic-conflicts2.doc.

POSNER D N (2000) Measuring Ethnic Identities and Attitudes Regarding Inter-group Relations: Methodological Pitfalls and a New Technique. Paper presented to LiCEP workshop, University of Pennsylvania, October 20-22,http://www.duke.edu/web/licep/2/posner/posner.pdf.

SILBER L and Little A (1997) Yugoslavia: Death of a Nation (revised and updated edition). New York: Penguin Books USA Inc.

SZAYNA T A (Ed.) (2000) Identifying Potential for Ethnic Conflict: Application of a Process Model. Santa Monica: RAND.

TURCOTTE D L (1999) Self-Organized Criticality. Reports on Progress in Physics 62, pp. 1377-1429.

WOODWARD S L (1995) Balkan Tragedy: Chaos and Dissolution after the Cold War. Washington D.C.: The Brookings Institution.


ButtonReturn to Contents of this issue

© Copyright Journal of Artificial Societies and Social Simulation, [2003]